V11 launch incident (2026-09-29 18:18 remote time)

All six training subprocesses stopped at destination.mkdir(exist_ok=False) with FileNotFoundError: the parent checkpoints/ directory had not been created at launch (earlier experiments created it in the launch command). No training step ran, no TRAIN, DEV or EVAL-V3 media was read and no metric was computed.

The six attempts are archived here by the runner's own archive_attempt (incident kind launch_missing_checkpoints_directory, recorded in audit/incidents.json) and count against the frozen infrastructure retry budget. The identical frozen commands are rerun after creating the directory. No frozen file, code or configuration changed.